Model-Based Deep Learning For Sensor Selection

Deep learning (DL) has emerged as a ubiquitous tool, tackling diverse challenges from distinguishing between cats and dogs to autonomous driving. Its efficacy often hinges on vast training datasets, yet concerns persist regarding interpretability, particularly in applications critical to human safety. Furthermore, labeled data scarcity complicates its application. In contrast, conventional signal and image processing methods leverage physical principles without necessitating extensive data, albeit requiring iterative convergence over numerous cycles.

This talk explores an alternative approach bridging these extremes: unrolling-based learning. This method constructs a neural network from iterative algorithms, combining the advantages of reduced data dependency, accelerated convergence, accuracy, and, crucially, interpretability. To illustrate these concepts practically, we delve into the sensor selection problem. Here, the objective is to identify a subset of sensors from a large pool. We discuss the significance of this problem, methodologies for resolution, and the application of model-based learning approaches.